Qwen3.6 27B on a 64 GB Mac: speed, quality, memory
Alibaba Qwen, released Apr 2026. 27B parameters, dense. 4-bit MLX build, 16.1 GB download. License Apache-2.0.
- HumanEval+ pass@1: 153 of 164 (93.3%), rank 1 of 8
- HumanEval pass@1: 158 of 164 (96.3%)
- Generation speed: 17.6 tokens per second (5 trials, 17.3 to 17.8); prompt 459 tokens per second
- Peak memory: 17.0 GB · power while generating: 24 W · GPU peak 80 °C
Run it yourself
pip install mlx-lm mlx_lm.chat --model mlx-community/Qwen3.6-27B-4bit
HumanEval/67
Read a string like '5 apples and 6 oranges' and work out how many mangoes are left.
It took the third word as the orange count. That word is 'and', so int('and') crashed. All 7 other models solved this one.
Problems it failed
HumanEval/67, HumanEval/92, HumanEval/116, HumanEval/134, HumanEval/145, HumanEval/163
Raw files: speed.log, evaluate.log, eval_results.json